Light Robotics is an AI robotics company built around compact foundation models for autonomous navigation and industrial automation. Light Robotics needed to modernize GPU-intensive model training and evaluation workflows without slowing active R&D. ASCENDING helped the team migrate from Azure to AWS and implement a scalable MLOps foundation with Amazon SageMaker and Amazon Bedrock.
BackgroundGPU-Heavy Robotics AI Development Outgrowing Azure
Light Robotics is a next-generation AI robotics company founded by a former OpenAI engineer, with teams operating across Singapore and China. As its compact AI models for robotics expanded into autonomous navigation, smart factories, and industrial automation use cases, the company needed faster iteration loops and more reliable model operations.
Its existing Azure-based training environment was becoming a bottleneck. The organization required a cloud-native platform that could support rapid experimentation, reproducible model training, and secure model evaluation interfaces for distributed teams.
The ChallengeAzure Bottlenecks Limiting Scale and Observability
The challenge is coordinating a multi-region MLOps migration without interrupting active robotics R&D. Light Robotics faced both technical and operating constraints while trying to scale model development.
- GPU-heavy model training on Azure was limiting scalability and cost efficiency.
- Training and deployment pipelines were manual, fragmented, and had limited observability.
- Cross-team evaluation workflows were difficult to coordinate across regions.
- Internal teams needed secure, low-latency access to foundation-model interfaces for robotics inference testing.
- Migration had to occur without disrupting in-flight R&D and prototyping timelines.
AWS Modernization and GenAI Platform Expertise
ASCENDING is the delivery partner Light Robotics engaged to modernize its ML platform without slowing R&D. ASCENDING was selected based on its AWS modernization experience and hands-on expertise in GenAI platform engineering. The team combined SageMaker performance optimization, Bedrock integration, and practical MLOps delivery to help Light Robotics replatform quickly while controlling risk and cost.
As an AWS Advanced Consulting Partner, ASCENDING aligned architecture decisions to measurable outcomes: faster training cycles, lower compute spend, and a repeatable path for future model operations.
The SolutionMLOps Transformation Built on SageMaker and Bedrock
The solution is a re-platformed MLOps stack that pairs managed training orchestration with secure model evaluation. ASCENDING designed and implemented an end-to-end ML transformation centered on Amazon SageMaker for training orchestration and Amazon Bedrock for secure model evaluation.
- Migrated GPU training workloads from Azure to Amazon SageMaker to improve scalability and operational control.
- Applied distributed training with Managed Spot Instances to reduce cost while maintaining performance.
- Implemented SageMaker Experiments for hyperparameter tracking, model versioning, and auditability.
- Built SageMaker Pipelines to automate end-to-end training orchestration and reproducibility.
- Exposed internal foundation models through Amazon Bedrock APIs for secure, prompt-driven evaluation workflows.
- Enabled cross-region model testing and prototype integration for teams in Singapore and China.
Faster Training, Lower Costs, Scalable MLOps Foundation
The outcome is a faster, lower-cost ML platform that scales with Light Robotics' commercial roadmap. The new AWS-based ML platform improved both development velocity and operational maturity for Light Robotics.
- Reduced training cycle time by 50%.
- Lowered compute cost by 30% through Spot-backed training optimization.
- Established fully automated, reproducible model training pipelines.
- Improved cross-region collaboration with hosted model evaluation interfaces for distributed research teams.
- Created a scalable foundation for commercial-grade robotics AI expansion.
Built with Amazon SageMaker and Bedrock MLOps Services
Frequently Asked Questions
How long does an Azure-to-AWS MLOps migration like this typically take?
Timeline depends on model complexity and pipeline maturity, but most GPU training migrations to Amazon SageMaker move in phases — piloting one training job before cutting over full production pipelines — so active R&D is never paused for a hard cutover.
How is model evaluation kept secure across distributed robotics teams?
Security is enforced through Amazon Bedrock's managed API layer, which lets Singapore- and China-based teams query foundation models for robotics inference testing without exposing underlying model weights or infrastructure.
Does training on Managed Spot Instances risk losing in-progress robotics model runs?
Interruption risk is mitigated with checkpointing and Spot interruption handling, and combined with Amazon CloudWatch monitoring, teams get visibility into training health while still capturing significant compute savings over on-demand GPU pricing.
